Papers by Junyang Wang
Turning the Tide: Repository-based Code Reflection (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Code large language models (LLMs) enhance programming by understanding and generating code across languages. |
| Approach: | a new benchmark evaluates code understanding and generation in repositories using code large language models. |
| Outcome: | The proposed model improves code understanding and generation in repositories by evaluating 1,888 test cases across 6 programming languages. |
Table-as-Search: Agentic Information Seeking is Table Completion (2026.findings-acl)
Copied to clipboard
Tian Lan, Felix Henry, Bin Zhu, Qianghuai Jia, Junyang Ren, Qihang PU, Haijun Li, Longyue Wang, Zhao Xu, Weihua Luo
| Challenge: | Current Information Seeking (InfoSeeking) agents struggle to maintain focus and coherence during long-horizon exploration, as tracking search states within one plain-text context is inherently fragile. |
| Approach: | They propose a structured planning framework that reformulates the InfoSeeking task as a Table Completion task. |
| Outcome: | The proposed framework outperforms state-of-the-art frameworks across three kinds of benchmarks, including multi-agent framework and commercial systems. |
Confidence v.s. Critique: A Decomposition of Self-Correction Capability for LLMs (2025.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to improve self-correction performance of Large Language Models are based on intrinsic selfcorrectione, which allows the model to check and revise its selfgenerated answers without external feedback. |
| Approach: | They propose to decompose the self-correction capability into confidence and critique capabilities and a metric for overall self-corretion capability evaluation. |
| Outcome: | The proposed method outperforms vanilla SFT and achieves much higher accuracy after self-correction. |
Outcome Accuracy is Not Enough: Aligning the Reasoning Process of Reward Models (2026.acl-long)
Copied to clipboard
Binghai Wang, Yantao Liu, Yuxuan Liu, Tianyi Tang, Shenzhi Wang, Chang Gao, Chujie Zheng, Yichang Zhang, Le Yu, Shixuan Liu, Tao Gui, Qi Zhang, Xuanjing Huang, Bowen Yu, Fei Huang, Junyang Lin
| Challenge: | Recent studies observe a phenomenon where reward models achieve high accuracy on static datasets but fail to generalize effectively during RLHF. |
| Approach: | They propose a method that combines rationale consistency with outcome accuracy to improve performance on RM-Bench and JudgeBench. |
| Outcome: | The proposed method surpasses baselines on RM-Bench and JudgeBench by an average of 5% and improves creative writing tasks by 7%. |
PLAWBENCH: A Rubric-Based Benchmark for Evaluating LLMs in Real-World Legal Practice (2026.acl-long)
Copied to clipboard
Yuzhen Shi, Huanghai Liu, Yiran HU, Song Gaojie, Xu Xinran, Yubo Ma, Tianyi Tang, Li Zhang, Qingjing Chen, Feng Di, Wenbo Lv, Weiheng Wu, Kexin Yang, Sen Yang, Wei Wang, Rongyao Shi, Qiu Yuanyang, Yuemeng Qi, Zhang Jingwen, Sui Xiaoyu, Yifan Chen, Zhang Yi, An Yang, Bowen Yu, Dayiheng Liu, Junyang Lin, Weixing Shen, Bing Zhao, Charles L. A. Clarke, HU Wei
| Challenge: | Existing benchmarks for large language models (LLMs) are coarse, single-dimensional metrics and do not explicitly assess fine-grained legal reasoning. |
| Approach: | They propose a Practical Law Benchmark to evaluate large language models in real-world legal practice scenarios. |
| Outcome: | The proposed model is based on 850 questions and 13 scenarios with expert-designed evaluation rubrics. |
Demons in the Detail: On Implementing Load Balancing Loss for Training Specialized Mixture-of-Expert Models (2025.acl-long)
Copied to clipboard
Zihan Qiu, Zeyu Huang, Bo Zheng, Kaiyue Wen, Zekun Wang, Rui Men, Ivan Titov, Dayiheng Liu, Jingren Zhou, Junyang Lin
| Challenge: | Existing Mixture-of-Experts training frameworks use a micro-batch to calculate LBL . micro-batches are restricted to a single sequence, preventing expert specialization . |
| Approach: | They propose to use a global-batch to loosen the load balance constraint for MoEs models . they propose to synchronize fi across micro-batches and then use it to calculate the LBL . |
| Outcome: | The proposed global-batch LBL improves the domain specialization of experts . the micro-battery LBL is almost at the sequence level, and the router is pushed to distribute the token evenly . |
START: Self-taught Reasoner with Tools (2025.emnlp-main)
Copied to clipboard
Chengpeng Li, Mingfeng Xue, Zhenru Zhang, Jiaxi Yang, Beichen Zhang, Bowen Yu, Binyuan Hui, Junyang Lin, Xiang Wang, Dayiheng Liu
| Challenge: | Large Reasoning Models (LRMs) have demonstrated remarkable capabilities in complex reasoning through long chain-of-thought, yet they struggle with precise computations and algorithmic operations. |
| Approach: | They propose a training-free approach that activates LRMs’ latent tool-use capabilities through artificial hints and a framework that enables models to learn effective tool utilization through diverse hint patterns and rejection-based data synthesis. |
| Outcome: | Experiments show that START significantly improves state-of-the-art LRMs across challenging benchmarks, including competition-level mathematics (AMC23: 95.0%, AIME24: 75.6%) and graduate-level science questions (GPQA: 64.6%). |
Sketch and Refine: Towards Faithful and Informative Table-to-Text Generation (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for table-to-text generation suffer from poor faithfulness and low coverage. |
| Approach: | They propose a method that combines Autoregressive and Non-Autoregressive generation to generate a table-to-text from a key-value table using a skeleton and an edit-based non-autoregressively generation model. |
| Outcome: | The proposed method outperforms the existing methods on WikiPerson and WikiBio datasets on coverage and faithfulness. |
HellaSwag-Pro: A Large-Scale Bilingual Benchmark for Evaluating the Robustness of LLMs in Commonsense Reasoning (2025.findings-acl)
Copied to clipboard
Xiaoyuan Li, Moxin Li, Rui Men, Yichang Zhang, Keqin Bao, Wenjie Wang, Fuli Feng, Dayiheng Liu, Junyang Lin
| Challenge: | Existing studies show that large language models are robust in commonsense reasoning . however, some variations in questions can lead to incorrect responses . |
| Approach: | They propose a large-scale bilingual benchmark consisting of 11,200 cases . they conduct extensive experiments on 41 representative LLMs . |
| Outcome: | The proposed benchmark systematically evaluates the robustness of large language models in commonsense reasoning. |
Parsing Natural Language into Propositional and First-Order Logic with Dual Reinforcement Learning (2022.coling-1)
Copied to clipboard
Xuantao Lu, Jingping Liu, Zhouhong Gu, Hanwen Tong, Chenhao Xie, Junyang Huang, Yanghua Xiao, Wenguang Wang
| Challenge: | Existing methods to parse natural language into structured logical expressions have limitations due to paucity of labeled data. |
| Approach: | They propose a scoring model to automatically learn a model-based reward . they also propose introducing a Chinese-PL/FOL dataset to compensate for paucity of labeled data . |
| Outcome: | The proposed model outperforms competitors on several datasets. |
Bag-of-Words as Target for Neural Machine Translation (P18-2)
Copied to clipboard
| Challenge: | Existing neural machine translation models only use one correct sentence as the target, and the other correct sentences are punished as the incorrect ones. |
| Approach: | They propose an approach that uses both the sentences and the bag-of-words as targets in the training stage to encourage the model to generate the potentially correct sentences that are not appeared in the train set. |
| Outcome: | The proposed model outperforms baseline models on a Chinese-English translation dataset by the BLEU score of 4.55. |
Prompt Tuning for Unified Multimodal Pretrained Models (2023.findings-acl)
Copied to clipboard
| Challenge: | Prompt tuning has demonstrated success in natural language pretraining and even vision pretraining. |
| Approach: | They propose to apply prompt tuning to a unified sequence-to-sequence pretrained model by adding a sequence of learnable embeddings to each layer and finetuning the pretrained models on downstream tasks. |
| Outcome: | The proposed method outperforms other parameter-efficient tuning methods on multimodal models and is robust against adversarial attacks. |
A Probabilistic Inference Scaling Theory for LLM Self-Correction (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated the capability to refine their generated answers through self-correction, enabling continuous performance improvement over multiple rounds. |
| Approach: | They propose a probabilistic theory to model the dynamics of accuracy change and explain performance improvements observed in multi-round self-correction. |
| Outcome: | The proposed model can predict accuracy curves and improve accuracy over multiple rounds. |
Rationales Are Not Silver Bullets: Measuring the Impact of Rationales on Model Performance and Reliability (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing studies have shown that training language models with rationales augmentation is beneficial, but this view does not hold consistently. |
| Approach: | They conduct comprehensive investigations to thoroughly inspect the impact of rationales on model performance and a novel perspective of model reliability. |
| Outcome: | The proposed method outperforms untrained models in several areas and provides informative regulations on the broad utilization of rationales. |
InSerter: Speech Instruction Following with Unsupervised Interleaved Pre-training (2025.acl-long)
Copied to clipboard
Dingdong Wang, Jin Xu, Ruihang Chu, Zhifang Guo, Xiong Wang, Jincenzi Wu, Dongchao Yang, Shengpeng Ji, Junyang Lin
| Challenge: | Recent advances in speech large language models exhibit suboptimal performance in adhering to speech instructions. |
| Approach: | They propose a method to pre-train large-scale unsupervised speech-text sequences . they use text-to-speech conversion to generate textual continuations corresponding to provided speech segments . |
| Outcome: | The proposed model achieves superior or competitive results across diverse speech processing tasks. |
RemoteRAG: A Privacy-Preserving LLM Cloud RAG Service (2025.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) have a tendency to generate factually incorrect or purely fictional responses, a phenomenon known as hallucination. |
| Approach: | They propose to use remote RAG to protect user query from privacy leakage . they introduce (n,)-DistanceDP to characterize privacy leakages of user query . |
| Outcome: | The proposed solution can resist embedding inversion attacks while achieving no loss in retrieval under various settings. |
Autoencoder as Assistant Supervisor: Improving Text Representation for Chinese Social Media Text Summarization (P18-2)
Copied to clipboard
| Challenge: | Existing abstractive text summarization models learn a semantic representation of the source text and the summaries from it. |
| Approach: | They evaluate the model on a popular Chinese social media dataset and compare it to other models. |
| Outcome: | The proposed model achieves state-of-the-art performance on a popular Chinese social media dataset. |
From Completion to Editing: Unlocking Context-Aware Code Infilling via Search-and-Replace Instruction Tuning (2026.acl-long)
Copied to clipboard
Jiajun Zhang, Zeyu Cui, Jiaxi Yang, Lei Zhang, Yuheng Jing, Zeyao Ma, Tianyi Bai, Zilei Wang, Qiang Liu, Liang Wang, Binyuan Hui, Junyang Lin
| Challenge: | Fill-in-the-Middle (FIM) models suffer from performance degradation and prohibitive latency. |
| Approach: | They propose a search-and-replace infilling framework that integrates agentic verification and editing into a single-pass inference process. |
| Outcome: | The proposed framework harmonizes completion tasks with the instruction-following priors of Chat LLMs, extending the paradigm from static infilling to dynamic context-aware editing. |
Transferring General Multimodal Pretrained Models to Text Recognition (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for text recognition rely on large-scale pretraining on human-annotated or synthetic data. |
| Approach: | They propose a method to transfer multimodal pretrained models to text recognition using image captioning. |
| Outcome: | The proposed method outperforms the baselines and achieves state-of-the-art performance in the Chinese text recognition benchmark. |
An Effective Deployment of Diffusion LM for Data Augmentation in Low-Resource Sentiment Classification (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing models for textual data augmentation (DA) are highly data-hungry and struggle to perform satisfactorily under noisy conditions. |
| Approach: | They propose to leverage a diffusion language model to capture in-domain knowledge and generate pseudo samples by reconstructing strong label-related tokens. |
| Outcome: | The proposed method captures in-domain knowledge and generates pseudo samples by reconstructing strong label-related tokens. |
SpeechLLM-as-Judges: Towards General and Interpretable Speech Quality Evaluation (2026.acl-long)
Copied to clipboard
Hui Wang, Jinghua Zhao, Yifan Yang, Shujie Liu, Junyang Chen, Yanzhe Zhang, Shiwan Zhao, Jinyu Li, Jiaming Zhou, Haoqin Sun, Yan Lu, Yong Qin
| Challenge: | Existing methods for evaluating the perceptual quality of synthetic speech are limited due to the complexity of perceptual quality factors and the diversity of speech generation tasks. |
| Approach: | They propose a new paradigm for enabling large language models to conduct structured speech quality evaluation using a large-scale dataset. |
| Outcome: | The proposed model performs well across tasks and languages. |
Imitation Learning for Non-Autoregressive Neural Machine Translation (P19-1)
Copied to clipboard
| Challenge: | Existing non-autoregressive translation models lack parallel decoding, which is a bottleneck for NMT decoding. |
| Approach: | They propose a framework for non-autoregressive machine translation that emulates the autoregressive model by sampling sentence length in parallel. |
| Outcome: | The proposed model achieves 31.85 BLEU on WMT16 RoEn and 30.68 BLUE on IWSLT16 EnDe on the IWSLD16, WMT14 and WMT15 datasets. |
LightMoE: Task-Aware Expert Availability Management for Memory-Efficient MoE-LLM Inference (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing solutions for balancing model accuracy with inference latency are limited due to memory constraints. |
| Approach: | They propose a framework for memory-efficient MoE inference that exploits the functional redundancy and temporal locality of expert activation. |
| Outcome: | The proposed framework improves accuracy-efficiency trade-off by 4.3% over pruned models and 2.4% over dynamic swapping methods while maintaining inference latency comparable to pruned model. |